Research for automation you can defend.

Experiments, papers, and field notes. Inspect the method and evidence before carrying a pattern into your next build.

Explore the research Browse experiments
Canon points a brass pencil at the extra peak on the second paper trace.
Concept illustration · Comparing two runs

From the research desk.

research2026-10-16

A Fresh Index Can Still Carry Old Facts

A synthetic Atlas/Substrate experiment on selective context: 20 frozen questions, exact citations, bounded responses, and the difference between refreshing a projection and verifying its sources.

research2026-06-24

Endpoint Construction Is Product Construction

Why AI-native products depend on the capability boundary more than the chat surface. A practical model for treating endpoints as the product grammar that defines intent, schema, authority, state, limits, errors, evidence, and fallback.

research2026-06-22

The Loop-Operable Codebase

A CREATE SOMETHING field paper on converting loop-engineering hype into a bounded repo-native operating system using Linear, Symphony, worktrees, legibility contracts, policy artifacts, and explicit dispatch gates.

research2026-06-22

The Proof Surface

Why agent work needs public receipts, private evidence, and owner authority once it leaves chat. A practical model for turning traces, deploys, approval notes, blocked states, delivery records, and workflow templates into business-readable proof.

research2026-06-22

The Eval Evidence Layer

How Langfuse traces and eval gates make agent workflows measurable. A quantitative model for turning Dify runtime traces, MCP eval gates, approval receipts, and blocked-state evidence into publish, hold, rollback, or graduation decisions.

research2026-06-21

The Policy OS Contract Bundle

Why governed AI workflows need a portable contract bundle before they need more autonomy. A practical model for defining MCP capability, agent behavior, outcome success, golden tasks, runbooks, and runtime graduation.

All 51 papers

Put the findings to work.

Inspect the references behind the work. Installation instructions and source stay close to the question they answer.

Browse plugins
Our working model

Database.
Automation.
Judgment.

Is the data available? Did the execution succeed? Was the right policy applied?

Read the framework ↗